How to minimise your eco impact while using AI

Maybe hearing about the staggering environmental cost of using generative AI has promoted you to read this article. Knowing how it works is the key to understanding why these two simple suggestions can significantly reduce your environmental impact when using Generative AI. Before others say it, early estimates of energy use are likely to be 10 times higher than reality and reasonable AI use is a small fraction of your entire energy budget if you live in a developed country. Regardless, there is still huge value in many people making small changes to reduce our burden on this groaning planet. So read on :)

LLMs have no memory. The model computes the probabilities of the next words (token) based on the words it is presented. All the words. From the start of your chat to the end. Each and every time you press enter. As you chat grows, the energy consumed to process each response increases. Each time you press "enter" in a chatbot, the entire chat history becomes the prompt - NOT just your latest interaction. Imagine this worse case: you have finished a long chat and, being polite, you say "thank you". Once you press enter, you have spend more energy at that moment that you have for every preceding interaction!

Eco-Strategy 1:

Consider what you are asking it to do and provide a long, explicit prompt with everything, rather than back and forth. Forget the illusion of a "conversation" that chatbots create - it is an illusion! Start a new chat regularly. If you have been using it to brainstorm (and you are not yet done), copy the top best 5 ideas you have converged on into a new chat and start again. This can save orders of magnitude of energy and hence eco-impact. There is an added bonus of this (which gives you a better result) - more on this later. I recommend that you start yourself a prompt library so that you can build and develop prompts, then cut and paste them, just changing the details. The more you give it (reducing the back and forth), the less energy you will use.

Diffusion models are the most common type of generativeAI used generate images. You know that LLMs use a lot of energy... multiply that by hundreds or thousands for one image! Here is how you can have both your image and lower your eco-impact.

Eco-Strategy 2:

Draft in text, only creating an image once you have a detailed prompt that perfectly matches what you are after. If anything is not explicit in your text prompt (like the fact that the leader in your image could be anything other than a white male) ... then spell it out in words. Here's an energy minimal approach to take:

1. Explicitly tell the AI you want text-only. Describe the idea you are imagining and ask the AI to brainstorm the best 6 ideas - in text only!

2. Choose the best idea[s] and ask it to develop a detailed text prompt for each feature (F.A.S.T. is a helpful mnemonic - what the Focus of the image; What are the elements and how are they Arranged; what Stylistic influences e.g. watercolour with pencil, photo realistic, etc ); What is the Tone of the image - lighting mood texture.)

3. Review your final text prompt - take it into a word processor to make final changes until you are happy.

4. Generate using your edited prompt. That will have saved you the useless images before you got it right - and the energy they represent!

Consider procedural image generation if you are after a graphic rather than an artwork. It uses way less energy, you have far more control and you can create animations and simulations - see this post.

There are so many added bonuses of a deeper understanding of how AI works. If you want to master generative AI (especially in an educational setting) check out my Professional Development course for teachers. -Taming the AI dragon, Module 1).

You might also enjoy other AI related posts you can find here.